How to Perform Object Detection With YOLOv3 in Keras

Last Updated on

Object detection is a task in computer vision that involves identifying the presence, location, and type of one or more objects in a given photograph.

It is a challenging problem that involves building upon methods for object recognition (e.g. where are they), object localization (e.g. what are their extent), and object classification (e.g. what are they).

In recent years, deep learning techniques are achieving state-of-the-art results for object detection, such as on standard benchmark datasets and in computer vision competitions. Notable is the “You Only Look Once,” or YOLO, family of Convolutional Neural Networks that achieve near state-of-the-art results with a single end-to-end model that can perform object detection in real-time.

In this tutorial, you will discover how to develop a YOLOv3 model for object detection on new photographs.

After completing this tutorial, you will know:

  • YOLO-based Convolutional Neural Network family of models for object detection and the most recent variation called YOLOv3.
  • The best-of-breed open source library implementation of the YOLOv3 for the Keras deep learning library.
  • How to use a pre-trained YOLOv3 to perform object localization and detection on new photographs.

Discover how to build models for photo classification, object detection, face recognition, and more in my new computer vision book, with 30 step-by-step tutorials and full source code.

Let’s get started.

  • Update Oct/2019: Updated and tested for Keras 2.3.0 API and TensorFlow 2.0.0.
How to Perform Object Detection With YOLOv3 in Keras

How to Perform Object Detection With YOLOv3 in Keras
Photo by David Berkowitz, some rights reserved.

Tutorial Overview

This tutorial is divided into three parts; they are:

  1. YOLO for Object Detection
  2. Experiencor YOLO3 Project
  3. Object Detection With YOLOv3

Want Results with Deep Learning for Computer Vision?

Take my free 7-day email crash course now (with sample code).

Click to sign-up and also get a free PDF Ebook version of the course.

Download Your FREE Mini-Course

YOLO for Object Detection

Object detection is a computer vision task that involves both localizing one or more objects within an image and classifying each object in the image.

It is a challenging computer vision task that requires both successful object localization in order to locate and draw a bounding box around each object in an image, and object classification to predict the correct class of object that was localized.

The “You Only Look Once,” or YOLO, family of models are a series of end-to-end deep learning models designed for fast object detection, developed by Joseph Redmon, et al. and first described in the 2015 paper titled “You Only Look Once: Unified, Real-Time Object Detection.”

The approach involves a single deep convolutional neural network (originally a version of GoogLeNet, later updated and called DarkNet based on VGG) that splits the input into a grid of cells and each cell directly predicts a bounding box and object classification. The result is a large number of candidate bounding boxes that are consolidated into a final prediction by a post-processing step.

There are three main variations of the approach, at the time of writing; they are YOLOv1, YOLOv2, and YOLOv3. The first version proposed the general architecture, whereas the second version refined the design and made use of predefined anchor boxes to improve bounding box proposal, and version three further refined the model architecture and training process.

Although the accuracy of the models is close but not as good as Region-Based Convolutional Neural Networks (R-CNNs), they are popular for object detection because of their detection speed, often demonstrated in real-time on video or with camera feed input.

A single neural network predicts bounding boxes and class probabilities directly from full images in one evaluation. Since the whole detection pipeline is a single network, it can be optimized end-to-end directly on detection performance.

You Only Look Once: Unified, Real-Time Object Detection, 2015.

In this tutorial, we will focus on using YOLOv3.

Experiencor YOLO3 for Keras Project

Source code for each version of YOLO is available, as well as pre-trained models.

The official DarkNet GitHub repository contains the source code for the YOLO versions mentioned in the papers, written in C. The repository provides a step-by-step tutorial on how to use the code for object detection.

It is a challenging model to implement from scratch, especially for beginners as it requires the development of many customized model elements for training and for prediction. For example, even using a pre-trained model directly requires sophisticated code to distill and interpret the predicted bounding boxes output by the model.

Instead of developing this code from scratch, we can use a third-party implementation. There are many third-party implementations designed for using YOLO with Keras, and none appear to be standardized and designed to be used as a library.

The YAD2K project was a de facto standard for YOLOv2 and provided scripts to convert the pre-trained weights into Keras format, use the pre-trained model to make predictions, and provided the code required to distill interpret the predicted bounding boxes. Many other third-party developers have used this code as a starting point and updated it to support YOLOv3.

Perhaps the most widely used project for using pre-trained the YOLO models is called “keras-yolo3: Training and Detecting Objects with YOLO3” by Huynh Ngoc Anh or experiencor. The code in the project has been made available under a permissive MIT open source license. Like YAD2K, it provides scripts to both load and use pre-trained YOLO models as well as transfer learning for developing YOLOv3 models on new datasets.

He also has a keras-yolo2 project that provides similar code for YOLOv2 as well as detailed tutorials on how to use the code in the repository. The keras-yolo3 project appears to be an updated version of that project.

Interestingly, experiencor has used the model as the basis for some experiments and trained versions of the YOLOv3 on standard object detection problems such as a kangaroo dataset, racoon dataset, red blood cell detection, and others. He has listed model performance, provided the model weights for download and provided YouTube videos of model behavior. For example:

We will use experiencor’s keras-yolo3 project as the basis for performing object detection with a YOLOv3 model in this tutorial.

In case the repository changes or is removed (which can happen with third-party open source projects), a fork of the code at the time of writing is provided.

Object Detection With YOLOv3

The keras-yolo3 project provides a lot of capability for using YOLOv3 models, including object detection, transfer learning, and training new models from scratch.

In this section, we will use a pre-trained model to perform object detection on an unseen photograph. This capability is available in a single Python file in the repository called “” that has about 435 lines. This script is, in fact, a program that will use pre-trained weights to prepare a model and use that model to perform object detection and output a model. It also depends upon OpenCV.

Instead of using this program directly, we will reuse elements from this program and develop our own scripts to first prepare and save a Keras YOLOv3 model, and then load the model to make a prediction for a new photograph.

Create and Save Model

The first step is to download the pre-trained model weights.

These were trained using the DarkNet code base on the MSCOCO dataset. Download the model weights and place them into your current working directory with the filename “yolov3.weights.” It is a large file and may take a moment to download depending on the speed of your internet connection.

Next, we need to define a Keras model that has the right number and type of layers to match the downloaded model weights. The model architecture is called a “DarkNet” and was originally loosely based on the VGG-16 model.

The “” script provides the make_yolov3_model() function to create the model for us, and the helper function _conv_block() that is used to create blocks of layers. These two functions can be copied directly from the script.

We can now define the Keras model for YOLOv3.

Next, we need to load the model weights. The model weights are stored in whatever format that was used by DarkNet. Rather than trying to decode the file manually, we can use the WeightReader class provided in the script.

To use the WeightReader, it is instantiated with the path to our weights file (e.g. ‘yolov3.weights‘). This will parse the file and load the model weights into memory in a format that we can set into our Keras model.

We can then call the load_weights() function of the WeightReader instance, passing in our defined Keras model to set the weights into the layers.

That’s it; we now have a YOLOv3 model for use.

We can save this model to a Keras compatible .h5 model file ready for later use.

We can tie all of this together; the complete code example including functions copied directly from the “” script is listed below.

Running the example may take a little less than one minute to execute on modern hardware.

As the weight file is loaded, you will see debug information reported about what was loaded, output by the WeightReader class.

At the end of the run, the model.h5 file is saved in your current working directory with approximately the same size as the original weight file (237MB), but ready to be loaded and used directly as a Keras model.

Make a Prediction

We need a new photo for object detection, ideally with objects that we know that the model knows about from the MSCOCO dataset.

We will use a photograph of three zebras taken by Boegh on safari, and released under a permissive license.

Photograph of Three Zebras

Photograph of Three Zebras
Taken by Boegh, some rights reserved.

Download the photograph and place it in your current working directory with the filename ‘zebra.jpg‘.

Making a prediction is straightforward, although interpreting the prediction requires some work.

The first step is to load the Keras model. This might be the slowest part of making a prediction.

Next, we need to load our new photograph and prepare it as suitable input to the model. The model expects inputs to be color images with the square shape of 416×416 pixels.

We can use the load_img() Keras function to load the image and the target_size argument to resize the image after loading. We can also use the img_to_array() function to convert the loaded PIL image object into a NumPy array, and then rescale the pixel values from 0-255 to 0-1 32-bit floating point values.

We will want to show the original photo again later, which means we will need to scale the bounding boxes of all detected objects from the square shape back to the original shape. As such, we can load the image and retrieve the original shape.

We can tie all of this together into a convenience function named load_image_pixels() that takes the filename and target size and returns the scaled pixel data ready to provide as input to the Keras model, as well as the original width and height of the image.

We can then call this function to load our photo of zebras.

We can now feed the photo into the Keras model and make a prediction.

That’s it, at least for making a prediction. The complete example is listed below.

Running the example returns a list of three NumPy arrays, the shape of which is displayed as output.

These arrays predict both the bounding boxes and class labels but are encoded. They must be interpreted.

Make a Prediction and Interpret Result

The output of the model is, in fact, encoded candidate bounding boxes from three different grid sizes, and the boxes are defined the context of anchor boxes, carefully chosen based on an analysis of the size of objects in the MSCOCO dataset.

The script provided by experiencor provides a function called decode_netout() that will take each one of the NumPy arrays, one at a time, and decode the candidate bounding boxes and class predictions. Further, any bounding boxes that don’t confidently describe an object (e.g. all class probabilities are below a threshold) are ignored. We will use a probability of 60% or 0.6. The function returns a list of BoundBox instances that define the corners of each bounding box in the context of the input image shape and class probabilities.

Next, the bounding boxes can be stretched back into the shape of the original image. This is helpful as it means that later we can plot the original image and draw the bounding boxes, hopefully detecting real objects.

The experiencor script provides the correct_yolo_boxes() function to perform this translation of bounding box coordinates, taking the list of bounding boxes, the original shape of our loaded photograph, and the shape of the input to the network as arguments. The coordinates of the bounding boxes are updated directly.

The model has predicted a lot of candidate bounding boxes, and most of the boxes will be referring to the same objects. The list of bounding boxes can be filtered and those boxes that overlap and refer to the same object can be merged. We can define the amount of overlap as a configuration parameter, in this case, 50% or 0.5. This filtering of bounding box regions is generally referred to as non-maximal suppression and is a required post-processing step.

The experiencor script provides this via the do_nms() function that takes the list of bounding boxes and a threshold parameter. Rather than purging the overlapping boxes, their predicted probability for their overlapping class is cleared. This allows the boxes to remain and be used if they also detect another object type.

This will leave us with the same number of boxes, but only very few of interest. We can retrieve just those boxes that strongly predict the presence of an object: that is are more than 60% confident. This can be achieved by enumerating over all boxes and checking the class prediction values. We can then look up the corresponding class label for the box and add it to the list. Each box must be considered for each class label, just in case the same box strongly predicts more than one object.

We can develop a get_boxes() function that does this and takes the list of boxes, known labels, and our classification threshold as arguments and returns parallel lists of boxes, labels, and scores.

We can call this function with our list of boxes.

We also need a list of strings containing the class labels known to the model in the correct order used during training, specifically those class labels from the MSCOCO dataset. Thankfully, this is provided in the experiencor script.

Now that we have those few boxes of strongly predicted objects, we can summarize them.

We can also plot our original photograph and draw the bounding box around each detected object. This can be achieved by retrieving the coordinates from each bounding box and creating a Rectangle object.

We can also draw a string with the class label and confidence.

The draw_boxes() function below implements this, taking the filename of the original photograph and the parallel lists of bounding boxes, labels and scores, and creates a plot showing all detected objects.

We can then call this function to plot our final result.

We now have all of the elements required to make a prediction using the YOLOv3 model, interpret the results, and plot them for review.

The full code listing, including the original and modified functions taken from the experiencor script, are listed below for completeness.

Running the example again prints the shape of the raw output from the model.

This is followed by a summary of the objects detected by the model and their confidence. We can see that the model has detected three zebra, all above 90% likelihood.

A plot of the photograph is created and the three bounding boxes are plotted. We can see that the model has indeed successfully detected the three zebra in the photograph.

Photograph of Three Zebra Each Detected with the YOLOv3 Model and Localized with Bounding Boxes

Photograph of Three Zebra Each Detected with the YOLOv3 Model and Localized with Bounding Boxes

Further Reading

This section provides more resources on the topic if you are looking to go deeper.




Other YOLO for Keras Projects


In this tutorial, you discovered how to develop a YOLOv3 model for object detection on new photographs.

Specifically, you learned:

  • YOLO-based Convolutional Neural Network family of models for object detection and the most recent variation called YOLOv3.
  • The best-of-breed open source library implementation of the YOLOv3 for the Keras deep learning library.
  • How to use a pre-trained YOLOv3 to perform object localization and detection on new photographs.

Do you have any questions?
Ask your questions in the comments below and I will do my best to answer.

Develop Deep Learning Models for Vision Today!

Deep Learning for Computer Vision

Develop Your Own Vision Models in Minutes

...with just a few lines of python code

Discover how in my new Ebook:
Deep Learning for Computer Vision

It provides self-study tutorials on topics like:
classification, object detection (yolo and rcnn), face recognition (vggface and facenet), data preparation and much more...

Finally Bring Deep Learning to your Vision Projects

Skip the Academics. Just Results.

See What's Inside

136 Responses to How to Perform Object Detection With YOLOv3 in Keras

  1. dong zhan May 28, 2019 at 7:48 am #

    thank you so much, machine learning for object detection! you have broadened my horizon, but, I am a game developer, so, usually bounding boxes are known, I have personal interest in go game, I wish I could understand machine-learning in go game, could you please give me a pointer?

    • Jason Brownlee May 28, 2019 at 8:23 am #


      Sorry, I don’t have any tutorials on games, I cannot give you good advice.

    • shgele October 2, 2019 at 5:35 am #

      Awsom @Jason Brownlee, I faced some issue like ValueError: If your data is in the form of symbolic tensors, you should specify the steps argument (instead of the batch_size argument, because symbolic tensors are expected to produce batches of input data) when I called Model.predict(image)

      • HEEGLS October 2, 2019 at 6:29 am #

        @Jasin Brownlee, It was very interesting and very well narrated. Could you please include , how would we add additional training set and labels ?
        Let say, I want this model to train on additional data, to classify Faces, or Hand Scribbled Digits and alphabets ? Suppose I have an additional training & test data set.

  2. Anna May 28, 2019 at 6:42 pm #

    Amazing post, thanks for sharing

  3. Salman Sajid May 28, 2019 at 11:19 pm #

    Kindly also make a blog on How we train a custom object detection with YOLO3

  4. Katia Z May 29, 2019 at 1:57 am #

    Jason, thank you very much for introducing computer vision so clearly! Could you specify in what part we saved our trained model to the disk?

    • Jason Brownlee May 29, 2019 at 8:50 am #

      The model was pre-trained, we simply loaded it.

      • Katia Z May 31, 2019 at 1:51 pm #

        Thank you so much for clean working code!) Could you advise some working and understandable GitHub repository for YOLOv3 for detecting objects on video, experiencer case is not clear in steps(

        • Jason Brownlee May 31, 2019 at 2:48 pm #

          I hope to cover this in the future, thanks for the suggestion.

  5. Noman Khan May 29, 2019 at 5:50 am #

    Thank you so much Sir! Very amazing and informative tutorial . I am beginner and following your tutorials for learning deep learning.
    kindly Sir!
    1) use YOLOv3 for camera video or simple video
    2) use LSTM for video sequence
    It will be great helpful for me

  6. Lina Xu June 1, 2019 at 1:13 am #

    Thanks so much for posting such a good tutorial, thanks!!!

  7. Jorge a June 1, 2019 at 4:31 pm #

    Thank you so much sir for your helpful tutorial.
    Could you tell me how can i change the number and titles of labels in this model ? or labels are unchangeable ?

  8. Lai Junhao June 4, 2019 at 12:54 am #

    Amazing guide Jason.

    At the start I experienced some difficulty with the library, since the latest version of Tensorflow did not work. I downgraded it to v1.12 and it worked.

    Thanks for the introduction to CV and YOLOv3. It was fun trying it out with my own images.

    • Jason Brownlee June 4, 2019 at 7:53 am #

      Thanks, and well done.

      I tested it with TensorFlow 1.13 on Ubuntu Linux and MacOS without incident.

      What OS/platform are you on?

      • Lai Junhao June 4, 2019 at 11:14 pm #

        i was using pycharm on windows using anaconda as the python interpreter.

        Irregardless, I will be moving on to the kangaroo tutorial since I managed to run the process successfully.

        Thanks again for creating these tutorials

      • Vikas Verma July 2, 2019 at 1:44 am #

        I am getting “AbortedError: Operation received an exception:Status: 3, message: could not create a dilated convolution forward descriptor, in file tensorflow/core/kernels/
        [[{{node conv_1/convolution}}]]” error in predictin step.

        Using tensorflow 1.13.1

  9. Niall O'Hara June 6, 2019 at 1:18 am #

    Great tutorial Jason. Very well laid out and easy to follow. I’m hoping to extend the tutorial to consider training on a custom data set. Would love to see a follow up blog on this πŸ™‚
    Keep up the great work!

  10. Hans Pfeufer June 14, 2019 at 7:47 pm #

    Great tutorial, thank you so much. How can I access the location of an object in an image? I would like to know, for instance, where the zebras are in the image (coordinates of the center of the bounding boxes) and save this data to a file.

    • Jason Brownlee June 15, 2019 at 6:31 am #

      The model will output bounding boxes (pixel coordinates) that you can use any way that you wish.

      In this tutorial, we simply draw the box.

  11. Ivan June 16, 2019 at 12:41 pm #

    Thanks for an ELI5 guide Jason!

    In interpreting the prediction array with decode_netout(), one of the argument was the anchors that you defined:

    anchors = [[116,90, 156,198, 373,326], [30,61, 62,45, 59,119], [10,13, 16,30, 33,23]]

    Is this suppose to be an initial guess for a bounding box of where the object may be?

    • Jason Brownlee June 17, 2019 at 8:13 am #

      Good question.

      It is the shape of the anchor boxes used during training. From the post:

      …and the boxes are defined the context of anchor boxes, carefully chosen based on an analysis of the size of objects in the MSCOCO dataset.

      • Paulo June 19, 2019 at 6:33 am #

        But what those anchors are? What are they used for? I’m guessing I my question is the same as Ivan’s.

        • Jason Brownlee June 19, 2019 at 8:19 am #

          They are chosen to best capture objects in the image net dataset, chosen based on that dataset.

          They are used when making a prediction to help quickly find objects in your new photos – e.g. bounding boxes in the image data.

          • Navin June 23, 2019 at 10:14 pm #

            Hi Jason,
            Question regarding anchors as they are hard coded, is it possible to derive them programmatically from yhat or the three NumPy array values or some other mechanism ?
            Also each array of ‘anchors’ having 6 elements is also puzzling, would be nice to know the process/documentation for defining these values.

            Good tutorial and helpful in getting started.
            A suggestion if you could cover on the the steps to generalize this code e.g. if I pass file path of a photo of a cat or a car etc. code should able to detect that just like it did for zebra ?

            Many thanks.

          • Jason Brownlee June 24, 2019 at 6:31 am #

            They are derived based on the average size of objects in a training dataset.

            You could derive them based on the expected object size in your dataset if you like.

  12. sumitra June 17, 2019 at 2:35 pm #

    Hi Jason,

    I have followed through the Experiencor YOLO3 for Keras Project to train with my own data set. If i would like to crop the bounding box from the newly predicted images, how do i go about it?

    Thank you

    • Jason Brownlee June 18, 2019 at 6:32 am #

      The model will output bounding boxes, you can use them directly on the image pixels.

      • sumitra June 18, 2019 at 2:58 pm #

        Thanks a lot for the response Jason. I managed to get good outputs with new images. However, the mAP performance of the model indicates 0.000 which is very strange. Could there be any reasons for it?

        • Jason Brownlee June 19, 2019 at 7:48 am #

          Well done.

          Perhaps test a suite of images to see if that uncovers more information?

  13. Chinmay June 25, 2019 at 3:51 pm #

    awesome tutorial sir !! Please answer this question
    many thanks

  14. Ashish Roopan June 27, 2019 at 2:20 am #

    Great work Jason.
    Can you also tell me how to train in with a new dataset?
    Mainly the format of annotation of the dataset to train with.(The YOLOv2 of this repository used .xml format like that of pascalVOC).

    Is it the same or something different?

    • Ashish Roopan June 27, 2019 at 3:40 am #

      Also I need train the model with transfer learning.
      So which all layers should I train?
      If possible can you please give an idea about the code to use?

      Thank you

    • Jason Brownlee June 27, 2019 at 7:58 am #

      I hope to cover that in the future.

      • xavier July 9, 2019 at 10:58 pm #

        Hi Jason, thx for this example !

        note that I found that there is some vertical down shift of the output boxes (it can be seen on your “zebra” image output above, the 3 boxes are sligthly too low. The shift can be really bigger on some other images I tested. Even on 416×416 input images)

        Any chance you fix this soon ?
        Many thanks anyway!

  15. Jorgetheman July 11, 2019 at 3:31 pm #


    This is an awesome tutorial. I went through it but at the end did not get a picture with bounding boxes as shown. I only got the array values and predictions for zebra and percentages.
    A figure with the bounding boxes wasn’t created.

    Any ideas?
    Thank you

  16. Karol Borkowski July 16, 2019 at 4:34 am #

    Is there in the book any additional content in addition to what is in this article?

  17. Karol Borkowski July 21, 2019 at 10:09 pm #

    Isn’t the compilation of the model missing?

  18. Gunners July 22, 2019 at 5:17 pm #

    I want to run this code. But it was too slow. I think it was not run on GPU. How could I test it by GPU? I want to run it in Jupyter Notebook.

  19. Mostafa July 24, 2019 at 6:18 am #

    Thanks so much for posting such a good tutorial. My question is how can we use YoloV3 to detect custom classes (labels)? In other words, how to reuse (retrain) it to detect custom objects?

  20. shreekanth July 24, 2019 at 8:15 pm #

    how can train the model with my data

  21. markos July 25, 2019 at 5:17 am #

    Excellent post, as always. One question. I have some yolo weights in .pt (pytorch) form. Is it possible to load this .pt file in some way or transform it so I can load it in a keras implemented yolo? Many thanks!!!

    • Jason Brownlee July 25, 2019 at 8:00 am #

      It may be possible to convert them.

      Sorry, I don’t have an example of this.

  22. markos July 27, 2019 at 7:09 pm #

    Thanks for your reply, Jason. I’ look it up.

  23. Karol Borkowski July 28, 2019 at 5:21 am #

    I have one technical question about the model. How the last layer knows which units correspond to which cell, since it is denselly connected with the previous one?

    • Jason Brownlee July 28, 2019 at 6:50 am #

      Sorry, what do you mean exactly?

      • Karol Borkowski July 28, 2019 at 8:33 am #

        The output of the YOLO network is S x S x N values, where S is the number of cells in both image directions. Therefore, the network predicts N values for each cell of the image.
        If I understand it correctly, each set on N values delivers a prediction for the corresponding image cell.
        My question is how a given image cell corresponds to the particular set of N values if everything is densely connected with everything? In other words, how the networks “know” which set of the predicted values are for which part of the image?

        I know that the question can be a liittle bit confusing, but I hope that now I have clarified it better.

  24. Karol Borkowski July 29, 2019 at 6:17 am #

    Let me ask about one more thing πŸ˜‰ The output is a tensor of size n x S x S x (B*5 + C), where n is the number of anchor sizes, S x S is the number of cells, B is the number of anchors for each size and C is the number of classess. In our model we have 3 anchors in 3 different sizes, and the MSCOCO dataset has 80 classes, so the shape of the output tensor should be: (3,S,S,3*5+80), what is (3,S,S,95). However, our output is (3,S,S,255). Why is that?

    • Jason Brownlee July 29, 2019 at 6:27 am #

      Not quite, the output is encoded.

      You can review the decoding functions in the post to see how the 3d output is interpreted.

      • Karol Borkowski July 29, 2019 at 6:37 am #

        Oh, I see it. Thanks again! πŸ˜‰

  25. Kumar July 31, 2019 at 2:45 am #

    I am new to machine learning. if we use pretrained weights for training custom (new) object, Will it detect old objects? for example, pretrained weights detect 80 objects (classes), I used this weight to train my new object (classes). Will it detect 81 objects or only one (new object).
    if it is detect only one (new object) how to make it detect 81 object.

    Thanks in advance

    • Jason Brownlee July 31, 2019 at 6:56 am #

      You will have to train on the a dataset that combines the original training dataset with your new class.

      Starting with pre-trained weights would speed this up a lot!

  26. FT August 2, 2019 at 5:12 pm #

    thanks a lot for this excellent post. I have a question. In the function correct_yolo_boxes(), what is the purpose to calculate offsets and scale? I thought offsets and scales are always 0 and 1 respectively, and they have no effect.

    • Jason Brownlee August 3, 2019 at 7:48 am #

      I believe we are scaling the boxes to the size of the image.

      • Madalin August 22, 2019 at 11:33 pm #

        Hi, is there a typo in the correct_yolo_boxes() implementation? The first line in the function: new_w, new_h = net_w, net_h will make the offsets and scales always 0 and 1.

        Thank you for your post, it helped me a lot

  27. Yashash Gaurav August 6, 2019 at 3:52 pm #

    Anyone trying to understand how the Yolov3 outputs are in the shapes they are, this( is an amazing resource.

    Thank you for the code walk through Jason!

  28. Prashanth August 7, 2019 at 4:44 pm #

    Excellent article, thanks.

    I trained my own YOLOv3 model with custom classes (fewer than the default 80) using Darknet. I therefore have a new .cfg file as well as weights file.
    I would like to perform inference using keras on Tensorflow, rather than Darknet.

    Is there a way I can do this using your program above? I don’t see a way to specify my own .cfg file. It only seems to accept the weights file.


    • Jason Brownlee August 8, 2019 at 6:29 am #

      There may be a way, I’m not sure off the cuff sorry. Some experimentation may be required.

      • Prashanth August 8, 2019 at 2:31 pm #

        I tried changing the number of filters from 255 to (num_classes + 5)*3, in each of the 3 YOLO layers. I’m able to get inference results, but they don’t match what I obtained with the original Darknet.

  29. Karthika August 20, 2019 at 1:27 pm #

    Thank you so much for such a simple and elaborate explanation.
    The code is working fine but I have a problem of not being able to save the model.
    I get the following error when I save any model.

    Error! F:\NITD\Project\Code\model.h5 is not UTF-8 encoded.
    Saving disabled.

    • Jason Brownlee August 20, 2019 at 2:13 pm #

      That is very odd.

      Perhaps it is an issue with your h5py library?

      Perhaps try posting/searching on stackoverflow for related issues?

  30. Sachin August 25, 2019 at 4:15 pm #

    Hi , I’m trying to perform this object detection on video, but i’m getting error . check this out.

    ERROR: rectangle() missing required argument ‘rec’ (pos 2)

    Here is the sample code. Do you know how to fix this? i tried a lot of things.

    def cv(frame, x, y, w, h, v_lables, v_scores):
    cv2.rectangle(img = frame,
    pt1 = (x, y),
    pt2 = (x + w, y + h),
    color = (0, 0, 255),
    thickness = 2,
    lineType = cv2.LINE_AA)
    cv2.putText(frame, “{}-{}”.format(v_labels,v_scores), (x + w, y),
    fontScale = 1, color = (255, 255, 255), thickness = 1, lineType = cv2.LINE_AA)
    return frame

    reader = imageio.get_reader(‘video.mp4’)
    fps = reader.get_meta_data()[‘fps’]
    writer = imageio.get_writer(‘output1.mp4’, fps = fps)

    for i,frame in enumerate(reader):
    image, image_w, image_h = load_image_pixels(frame, (input_w, input_h))
    yhat = model.predict(image)
    for j in range(len(yhat)):
    boxes += decode_netout(yhat[j][0], anchors[j], class_threshold, input_h, input_w)
    correct_yolo_boxes(boxes, image_h, image_w, input_h, input_w)
    do_nms(boxes, 0.5)
    v_boxes, v_labels, v_scores = get_boxes(boxes, labels, class_threshold)
    for z in range(len(v_boxes)):
    box = v_boxes[z]
    y1, x1, y2, x2 = box.ymin, box.xmin, box.ymax, box.xmax
    width = x2 – x1
    height = y2 – y1
    frame = cv(frame, x1, y1, width, height, v_labels[z], v_scores[z])

  31. Sachin August 26, 2019 at 3:22 pm #

    That’s okay i fixed it. Thank you !

    • Jason Brownlee August 27, 2019 at 6:24 am #

      I’m happy to hear that.

    • Karthika August 28, 2019 at 3:57 pm #

      Can you share the code ?

  32. Tedi August 27, 2019 at 1:02 am #

    Hello Jason,

    Thank you for the comprehensive tutorial.

    2 questions:

    1) I tried removing some of the objects in the labels and the image opened without applying the algorithm, there were no markings on it. How can we add objects such as (Jar, Speaker, etc..) or modify this current list?

    2) When the algorithm works, it does perform well. could we change the bounding box anchors from the code, or it has to be done from the training source code? In case we wanted to increase the height of the box.

    I understand these are in-depth questions, if you could provide any brief steps or research hints from your Books/Posts, that would be appreciated.


  33. Tedi August 28, 2019 at 7:36 am #

    Thank you!

    I will check it out.

  34. Pasin September 2, 2019 at 4:58 pm #

    So I tried to predict an image with the same code as written above but I got this error:

    ValueError: If your data is in the form of symbolic tensors, you should specify the steps argument (instead of the batch_size argument, because symbolic tensors are expected to produce batches of input data).

    Any idea on what’s wrong here? Thanks in advance!

    • Jason Brownlee September 3, 2019 at 6:13 am #

      Perhaps confirm that your libraries are up to date and that you loaded your image as a NumPy array?

  35. Hongbo Ai September 3, 2019 at 8:29 pm #

    the code works for zebra.jpg, but failed to the elephant and carrot pictures which are downloaded from internet

    • Jason Brownlee September 4, 2019 at 5:57 am #

      Perhaps double check hat the images were loaded correctly?

      • Hongbo Ai September 4, 2019 at 11:47 am #

        it is the threshold that filter out the correct result. anyway it works now.

        • Jason Brownlee September 4, 2019 at 1:43 pm #

          Happy to hear that.

        • Fetulhak September 14, 2019 at 9:55 pm #

          what do you mean exactly it is the threshold to filter the correct result? the same problem happened to me when i try using an elephant picture it did not give me the BB? Should I have to change the threshold value every-time I give another picture of different class?

          • Hongbo Ai September 23, 2019 at 7:42 pm #

            the picture outputs a class value as 0.7 while my threshlod is 0.8, so there is no output in the plot picture. when i change the threshlod to 0.6, the plot picture show the correct result

  36. Salman Sajd September 5, 2019 at 3:06 am #

    Thanks for a brilliant blog and have a one request
    Just like ”custom training on MaskR-CNN” blog also make a blog on YOLO to.

  37. Sridharan September 6, 2019 at 2:30 pm #

    Sir, how to store the objects detected in the bounding box as a separate image.

  38. Fetulhak September 14, 2019 at 9:33 pm #

    Thank you for your amazing dedication to teach those of us who are learning through the internet. the internet is my university. Currently I am working on project for detection of malaria parasite from thick blood microscopic images using yolov3 algorithm. what i am afraid of is may be yolov3 will not detect object at parasite level since they are tiny. what is your recommendation for me? I have the microscopic images.

    • Jason Brownlee September 15, 2019 at 6:21 am #

      That sounds like a great project.

      I would encourage you to test a suite of different models and discover what works best for your specific dataset.

  39. Bhaskar Chandra Trivedi September 18, 2019 at 4:54 am #

    Can we have tutorial for object tracking using discrimitive correlation filter

  40. Arelis September 30, 2019 at 10:53 am #

    Thank you. Your explain very good.
    Only question. How I can keep on disk the image with detection object.

    Good bless you

  41. Florin Andrei October 2, 2019 at 2:56 pm #

    Very cool example, but decode_netout() is very slow. Any suggestions for making it faster? Aggressively pruning unwanted items?

    E.g., what if I’m only interested in one type of label?

  42. Sandeco October 7, 2019 at 11:28 pm #

    Hi Jason, Thanks for all codes man. My friend, I updated tensorflow to 2.0 version, this code doesn’t work. I changed the importations to tensorflow.keras, but the class “Add” have a problem em this line “return add([skip_connection, x]) if skip else x”. Do u have some tip?

    • Jason Brownlee October 8, 2019 at 8:02 am #

      If you use standalone Keras v2.3.0 on top of TensorFlow 2.0, the code works fine.

  43. Dang Tien Dat October 8, 2019 at 6:34 pm #

    Thank you very much for sharing, I think I understand more about YOLOv3.
    But I still do not understand about meaning of numbers in :
    [(1, 13, 13, 255), (1, 26, 26, 255), (1, 52, 52, 255)]
    I think 13*13 or 26*26 and 52*52 is three different sizes of anchor boxes.
    But what is meaning of number ‘1’ and ‘255’ ?
    Let me thank you again for helping.

    • Jason Brownlee October 9, 2019 at 8:07 am #

      Good question.

      Yes, the first dimension (1) can be ignored, the middle are the size (13 or 26), and the final are the number of boxes, that will need to be reduced or interpreted.

  44. tyt October 16, 2019 at 8:31 pm #


    how can i make detection from this code

    • Jason Brownlee October 17, 2019 at 6:29 am #

      See the section titled “Make a Prediction”.

Leave a Reply